Increasing Cone-beam projection usage by temporal fitting

A Cone-beam CT system can be used to image the lung region. The system records 2D projections which will allow 3D reconstruction however a reconstruction based on all projections will lead to a blurred reconstruction in regions were respiratory motion occur. To avoid this the projections are typically positioned on the breathing cycle using the Amsterdam shroud method [7] or some external measurement device. Measurement with similar respiratory positions are grouped as belonging to the same respiration phase. This preprocessing is known as phase binning and allows for the reconstruction of each sorted data set. The common method of choice for reconstructing the 3D volume is the Feldkamp-Davis-Kress algorithm [2], however this method suffers from serious artefacts when the sample number of projections is too low which can happen due to phase binning. Iterative methods based on solving the forward projection problem [1] are known to be more robust in these situations. We study how the lower projection limits of an iterative method can be pushed even further by modelling a temporal relation between the respiratory phases. Although phase binned data is assumed the approach will work with raw measurements. It has been suggested in [8] to circumvent the Cone beam CT(CBCT) reconstruction by utilizing an ordinary planning CT instead and learning its deformation from the CBCT projection data. The main problem with this approach is that pathological changes can cause problems. Alternatively as suggested in [6] prior knowledge of the lung deformation estimated from the planning CT could be used to include all projections into the reconstruction. It has also been attempted to estimate both the motion and 3D volume simultaneously in [4]. Problems with motion estimation are ill-posed leading to suboptimal motion which in return affects the reconstruction. By directly including time into the image representation the effect of suboptimal motion fields are avoided and we are still capable of using phase neighbour projections. The 4D image model is fitted by solving a statistical cost function based on Poisons assumptions using an L-BFGS-B optimizer [5]. It will be demonstrated on a phantom data set that the information gained from a 4D model leads to smaller reconstruction errors than a 3D iterative reconstruction based on phase binned data.